MIT Students Built a Working Jet Engine With AI, Then Found Where AI Breaks

Can a general-purpose AI copilot help build a machine in which a small error can turn spinning metal and burning fuel into a serious hazard? MIT’s inaugural JARVIS Challenge offered an unusually concrete test: 31 undergraduates, divided into seven teams, were asked to conceive, fabricate and hot-fire a small single-spool gas-turbine engine while using frontier language models as engineering partners.

From prompts to a firing engine

The project’s target was not a presentation or a computer simulation. Teams aimed to build an engine running on Jet-A fuel, producing 50 to 100 pounds of thrust and completing repeated 60-second runs. Students used AI to summarize technical material, compare architectures, learn unfamiliar software, identify suppliers, organize calculations and even manage project tasks. They also had conventional engineering tools, MIT machine shops, test rigs and faculty safety oversight.

By the end of May, two senior teams had reached full-engine testing. One AI-assisted design ignited but suffered a rotor rub that seized the machine. The winning 811 Crew successfully started its engine, transitioned it to Jet-A and generated net thrust. That makes the result visually dramatic, but its real value is the comparison between what the models accelerated and what they could not reliably do.

AI was useful, but not an engineer

The models were strongest at information retrieval, explanation, trade studies and administrative work. They were much weaker when detailed geometry, physical intuition and safety-critical judgment became central. Students encountered hallucinated answers, excessive agreement and suggestions that lacked a grounded understanding of how parts would behave after fabrication. Some teams lost confidence in the tools after poor early answers.

Manufacturing, rather than design analysis, remained the main bottleneck. AI could suggest vendors but could not create trusted supplier relationships, shorten lead times or physically correct a misaligned component. The teams with deeper experience were better able to reject bad suggestions, while less experienced students sometimes used the tools more creatively. The apparent “sweet spot” was therefore neither blind adoption nor blanket skepticism: it was domain expertise combined with active, critical use.

Why it matters

Aerospace development is slow partly because design, procurement, fabrication and testing are tightly coupled. If qualified teams can use AI to compress portions of that cycle, smaller groups may explore more alternatives and reach hardware tests faster. The implications could extend beyond aviation to energy systems, robotics and other complex machinery.

But JARVIS was an educational challenge, not a controlled industrial trial. Seven student teams are too small a sample for broad productivity claims, sponsors provided unusually generous access to tools and facilities, and the project did not establish that AI-designed hardware is safer or better than conventionally engineered equipment. The winning team also entered with stronger relevant knowledge and used AI cautiously. The most defensible conclusion is narrower: today’s AI can multiply capable engineering work, but it cannot replace accountability, first-principles knowledge, testing or human hands.